Open-source spam call and text blocker for Android
15+ layer detection + Gradient-Boosted Tree ML | 32,933 spam numbers | Real-time caller ID | RCS filter | No API keys

CallShield blocks spam calls and texts using a 15+ layer on-device detection engine with a gradient-boosted tree ML scorer, campaign burst detection, RCS notification filter, and real-time caller ID overlay. Powered by a 32,933-number database with 30-minute hot list updates. Community-maintained, no accounts, no tracking.

• STIR/SHAKEN Trusted-Caller Allow — new detection layer that short-circuits heuristic / ML / campaign-burst when the carrier signs a PASSED attestation. Paired with the existing FAILED-blocks layer for a clean PASSED / FAILED / NOT_VERIFIED decision table. Priority slot sits below every explicit user rule — your blocklist always wins over the carrier.
• Auto-Mute Low-Confidence Blocks — opt-in setting that silences blocks scoring below 60% confidence to voicemail instead of hard-rejecting them, so you can review uncertain calls after the fact. High-confidence hits (database, blocklist, STIR fail, heuristic ≥ 60) still hard-reject even with auto-mute on.
• Centralized block-response decision table — buildBlockResponse() helper in CallShieldScreeningService with a pure shouldSilence() companion. Three silence/reject branches (silent-voicemail / auto-mute / hard-reject) now share one reviewable code path. Behavior-preserving refactor — existing response shapes emit identical CallResponse.Builder() calls.
• 14 new JVM unit tests — full priority-ladder regression sweep in StirShakenTrustCheckerTest and every branch of the silence/reject table in CallShieldScreeningServiceAutoMuteTest.
• 597 total unit tests + GitHub Actions CI — automated test pipeline on every push.
• Gradient-Boosted Tree ML model — 20 features, pure Kotlin, no TFLite dependency.
• Campaign burst detection — NPA-NXX prefix clustering identifies coordinated spam waves.
• Full accessibility — content descriptions across Compose UI, 48dp minimum touch targets.

1. 32,933 confirmed spam numbers — sourced from 1.75M FCC consumer complaints (2+ reports each), FTC Do Not Call, ToastedSpam, and community reports
2. 15+ layer detection + ML — database, heuristics, campaign burst detection, on-device gradient-boosted tree, SMS content analysis, RCS filter, STIR/SHAKEN, and more
3. Real-time caller ID overlay — parallel lookups against SkipCalls, PhoneBlock, WhoCalledMe + OpenCNAM caller name, with SIT tone anti-autodialer
4. 30-minute hot list — trending spam numbers and campaign ranges refresh every 30 minutes via GitHub Actions
5. Callback-aware — won't block callbacks from numbers you recently called, or urgent repeated callers
6. Community-driven — one-tap anonymous contribution via Cloudflare Worker, daily merge into database

All detection layers implement a shared IChecker interface and run in priority order via CheckerPipeline.run — first non-null result wins, every layer is testable in isolation. Priorities are stable numbers; the ladder below is the live order.

SMS-specific layers (append after the shared chain): SMS Context Trust → SMS Keyword Rules (with schedule) → SMS Content Analysis (30+ regex patterns, URL shorteners, suspicious TLDs, spam domain blocklist).

• Caller ID Overlay — suspicious calls (heuristic score 30-59) trigger a live multi-source lookup overlay with SkipCalls, PhoneBlock, WhoCalledMe + OpenCNAM caller name
• RCS Filter — NotificationListenerService monitors Google/Samsung Messages for RCS spam
• URL Safety — URLhaus (abuse.ch) checks for phishing/malware URLs in SMS/RCS (post-decision, notification only)
• STIR/SHAKEN — blocks calls failing carrier caller ID verification (Android 11+)
• After-Call Feedback — "Was this spam?" notification after suspicious calls for community reporting

Any wildcard, range, or SMS keyword rule can be time-gated to specific days of the week and an hour window. The hour picker supports overnight wrap; daysMask = 0 is the "no gating" sentinel so rules created before v1.6 behave identically.

When a call comes in, CallShield shows a real-time overlay that queries 4 sources simultaneously:

┌──────────────────────────────────┐
│ LIKELY SPAM                      │
│ (212) 555-1234                   │
│ New York, NY                     │
│ Spam Score: 80% (17 reports)     │
│ JOHN DOE (OpenCNAM)             │
│ ⚠ SkipCalls: Flagged            │
│ ⚠ PhoneBlock: 5 reports         │
│ ⚠ WhoCalledMe: 12 reports       │
│ All sources checked              │
│ [Search] [Block] [Dismiss]       │
│ 🔈 Play SIT Tone (anti-dialer)  │
└──────────────────────────────────┘

• Shows instantly with area code, then updates live as each source responds
• OpenCNAM caller name lookup (free, 60 req/hr)
• SIT Tone — ITU-T E.180 three-tone sequence tricks autodialers into removing your number
• Color-coded: green (safe) → yellow → orange → red (spam)

On-device 20-feature gradient-boosted tree model — pure Kotlin, no TFLite, no heavy ML libraries. Runs in microseconds.

Trained weekly from the CallShield database (50K positive + 50K negative samples). Threshold: 0.7 (conservative).

• Instant spam check through all 15+ detection layers with animated score gauge (0-100)
• Auto-paste from clipboard, area code lookup (330+ US/CA), haptic feedback
• Multi-source reverse lookup: SkipCalls + PhoneBlock + WhoCalledMe + OpenCNAM

• Recent calls with contact names, risk indicators, call type icons, filter chips (All/Missed/Spam)
• Blocked log with swipe-to-dismiss + undo, grouping with severity-scaled accent bars, filter chips
• Staggered entrance animations, shimmer loading skeletons

• Blocklist, Wildcards, Keywords, Whitelist, Database
• Export/import blocklists as JSON, per-rule enable/disable toggles
• Regex validation before adding wildcard rules

• Weekly bar chart with daily breakdown
• Detection source donut chart
• Monthly trend line
• Top offenders, area code heatmap, hourly heatmap

• Smart suggestions — detects area code spam patterns, one-tap block entire area code
• Weekly trend indicator — shows if spam is increasing or decreasing vs last week
• Last blocked preview card on dashboard with tap-to-inspect
• Blocking profiles: Work / Personal / Sleep / Maximum / Off
• Callback detection + repeated urgent caller allow-through
• FTC Do Not Call complaint filing
• After-call "Was this spam?" feedback notification

• Today vs yesterday blocked count with trend indicator
• Last blocked number and time
• Quick-access to lookup and protection status

• One-tap anonymous contribution via Cloudflare Worker
• False positive reporting subtracts votes
• Share spam warnings to any app

• Full backup/restore, CSV log export, auto-cleanup (7/14/30/90 days)
• Weekly full sync + 30-minute hot list refresh, daily digest notification
• Quick Settings tile, app shortcuts, home screen widget
• Protection test validates all layers and permissions
• Onboarding wizard with permission requests

• Network security config — cleartext traffic disabled in production
• Signing credentials — stored in local.properties, not hardcoded in build files
• Restricted FileProvider paths — scoped to export directory only
• Database-only backup — android:fullBackupContent excludes preferences containing API keys

All detection runs on-device. No personal data is collected. Network requests:

• Syncing spam database from GitHub (public)
• Real-time lookups against free public APIs (number queried, not stored)
• Community reports to Cloudflare Worker (phone number only, no identity)
• URLhaus checks for SMS URL safety (URL only)

No API keys required. No accounts. No analytics. No ads.

• Android 10+ (API 29)
• STIR/SHAKEN requires Android 11+ (API 30)
• Caller ID overlay requires "Display over other apps" permission
• RCS filter requires Notification Access permission

./gradlew assembleRelease

Requires JDK 17+. Signed APK at app/build/outputs/apk/release/app-release.apk.

Signing: Create local.properties in the project root with your keystore credentials:

RELEASE_STORE_FILE=path/to/keystore.jks
RELEASE_STORE_PASSWORD=...
RELEASE_KEY_ALIAS=...
RELEASE_KEY_PASSWORD=...

./gradlew testDebugUnitTest   # 210 tests

CI runs automatically via GitHub Actions on every push and pull request.

For deep technical details, see CLAUDE.md.

MIT